The transmission control protocol/internet protocol (TCP/IP) is a protocol that has been widely utilized for communications. Conventional network interface cards (NICs) typically contain specialized processors or accelerators that may be adapted to handle the processing of packetized information received from a transmission medium. In a typical network interface card, the reception of data may include processing of packetized data in a plurality of communications layers before the data is copied to its final destination, for example, an application buffer. However, receiving, buffering, processing and storing the packetized data communicated in TCP segments can consume a substantial amount of host processing power and memory bandwidth at the receiver. With today's high speed communication systems of the order of Gigabits, these conventional network interface cards are inefficient and unable to manage such high speeds.
TCP segmentation is a technology that may permit a very small portion of TCP processing to be offloaded to a network interface card (NIC). In this regard, a NIC that supports TCP segmentation does not truly incorporate a full transmission control processing offload engine. Rather, a NIC that supports TCP segmentation only has the capability to segment outbound TCP blocks into packets having a size equivalent to that which the physical medium supports. Each of the outbound TCP blocks is smaller than a permissible TCP window size. For example, an Ethernet network interface card that supports TCP Segmentation, may segment a 4 KB block of TCP data into 3 Ethernet packets. The maximum size of an Ethernet packet is 1518 bytes inclusive of header and a trailing CRC.
A device that supports TCP segmentation does track certain TOP state information such as the TCP sequence number that is related to the data that the offload NIC is segmenting. However, the device that supports TCP segmentation does not track any state information that is related to inbound traffic, or any state information that is required to support TCP acknowledgements or flow control. A NIC that supports full TCP offload in the established state is responsible for handling TCP flow control, and responsible for handling incoming TCP acknowledgements, and generating outbound TCP acknowledgements for incoming data.
TCP segmentation may be viewed as a subset of TCP offload. TCP segmentation allows the protocol stack or operating system to pass information in the form of blocks of TCP data that has not been segmented into individual TCP packets to a device driver. The block of data may be 4 Kbytes or 16 Kbytes. A network adapter associated with the device driver may acquire the blocks of TCP data, packetize the acquired blocks of TCP data into 1518-byte Ethernet packets and update certain fields in each incrementally created packet. For example, the network adapter may update a corresponding TCP sequence number for each of the TCP packets by incrementing the TCP sequence number for each of the packets. In another example, an IP identification (IP ID) field and flag field would also have to be updated for each packet. One limitation with TCP segmentation is that TCP segmentation may only be done on a block of data that is less than a TCP window size. This is due to the fact that a device implementing TCP segmentation has no influence over TCP flow control. Accordingly, the device implementing TCP flow control only segments outbound TCP packets.
A TCP segmentation device does not examine incoming packets and as such, has no influence over flow control. Any received acknowledgement packet is passed up to the host for processing. In this regard, acknowledgement packets that are utilized for flow control are not processed by the TCP segmentation device. Moreover, a TCP segmentation device does not perform congestion control or flow startup and does not calculate or modify any variables that are passed back to the operating system and/or host system processor.
Another limitation with TCP segmentation is that information tracked by TCP segmentation is only information that is pertinent for the lifetime of the TCP data. In this regard, for example, the TCP segmentation device may track TCP segmentation numbers but not TCP acknowledgement (ACK) numbers. Accordingly, the TCP segmentation device tracks only a minimal subset of information related to corresponding TCP data. This limits the capability and/or functionality of the TCP segmentation device. A further limitation with TCP segmentation is that a TCP segmentation device does not pass TCP processed information back to an operating system and/or host processor. This lack of feedback limits the TCP processing that otherwise may be achieved by an operating system and/or host system processor.
Other limitations associated with TCP segmentation are set forth in U.S. patent application Ser. No. 10/652,183, filed Aug. 29, 2003, which is incorporated herein by reference in its entirety.
Since the processing of TCP segments may consume a substantial amount of host processing power and memory bandwidth, in order to alleviate consumption of host resources, some of the TCP processing may be offloaded from the host as shown in FIG. 1. FIG. 1 illustrates a conventional offload system. Referring to FIG. 1, the system may include a CPU 10, a memory controller 20, a host memory 30, a host interface 40, a network interface card (NIC) 50 and an Ethernet 60. The NIC 50 includes a TCP offload engine (TOE) 70, a transmission frame buffer 80 and a reception frame buffer 90. The CPU 10 is coupled to the memory controller 20. The memory controller 20 is coupled to the host memory 30 and to the host interface 40. The host interface 40 is coupled to the NIC 50 via the TOE 70. The TOE 70 is coupled to the transmission frame buffer 80, the reception frame buffer 90 and the Ethernet 60.
In operation, incoming frames from the Ethernet 60 are received by the NIC 50. The TOE 70 processes the frames and stores them in the reception frame buffer 90. When buffers are available in the host memory 30 and when sufficient frames have been stored, the TOE 70 receives the frames stored in the reception buffer 90 and sends the frames to host memory 30 via the host interface 40 and the memory controller 20. Outgoing frames from the host are sent to the TOE 70 which stores them in the transmission frame buffer 80. When transmitting, the TOE 70 retrieves the frames stored in the transmission frame buffer 80 and transmits them via the Ethernet 60. For high-speed networking such as 10 Gigabits per second Ethernet (GbE), additional copying of data may add unnecessary strain on a computer's or host's memory sub-system. The memory subsystem of most commercially available servers or host computers becomes a bottleneck, thereby preventing the system from supporting high data rates such as 10 Gigabit network traffic. Since TCP/IP is the dominant transport protocol utilized by most applications today, it would therefore be useful to ease the burden of this processing to achieve, for example, scalable low CPU utilization when communicating with a peer machine.
TCP/IP utilizes a datagram service at the IP layer. Under normal operational conditions with router or switch congestion, IP datagrams may be dropped, leading to a “hole” in the stream of datagrams that are on their way to the receiver. The receiver may therefore receive datagrams out of order. Packet drop may also be the result of, for example, other less frequent transmission errors. The common way to deal with this is to buffer the datagrams that were successfully received, while waiting to get the missing datagram or datagrams by retransmission from the source. Retransmission may be triggered by the sender or the receiver. The TCP protocol allows a complete TCP Window of datagrams per connection to be on-flight from the sender to the receiver assuming a high performance configuration. The datagrams may contain 64 KBytes of data, for example. Many applications employ a large number of TCP connections, for example, 1000 to 100,000 TCP connections, to be supported by the receiver. At higher network speeds such as 1 Gigabit per second and higher, it would be inefficient to discard or drain the pipe or a portion of a received data stream every time there is a dropped datagram. TCP bandwidth probing methodologies such as slow start and/or congestion avoidance, which may be triggered at connection startup or when congestion is detected, may result in the loss of precious time and is inefficient since the congestion window is decreased and has to be gradually increased until it is equivalent to a receiver's advertised window size. Therefore, typical TCP implementations set aside a large buffer such as 64 MB to 6.4 GB to handle these situations. This large buffer is used to reassemble TCP/IP data, or IP fragments. The depth of the buffer may be dependent upon the product of connection bandwidth and network delay on the TCP connection. This architecture is therefore sensitive to LAN or WAN configuration and in this regard, more buffers may be utilized for a medium bandwidth, high-delay WAN configuration than for a low delay, high-speed LAN configuration.
The TCP offload architecture illustrated in FIG. 1 is also known as a store & forward approach. It adds latencies that are utilized to store the data in the buffers 80, 90 of the NIC 50, to manage the buffers 80, 90 and to retrieve information in an ordered fashion out of the buffers 80, 90 and into the host memory 30. During reception, received packets may be stored in the receive frame buffer 90 where they are processed. When packets arrive out-of-sequence, instead of dropping previously received associated packets, the received packets are buffered until missing packets are subsequently received. The received missing packets and the out-of-sequence packets are then reassembled or reordered. The assembled or reordered packets are then processed to determine where they should be placed on the host system. Once the placement of the assembled packets is determined, the assembled packets are then passed to the host where they are stored for processing. This respective buffering, processing, reassembling or reordering, processing and placement requires an excessive amount of memory and consumes an extensive amount of processing resources.
Similar considerations are applicable for the transmit side. A TCP sender maintains a transmit frame buffer 80 with all the data it has transmitted as part of the TCP “window”. Once the remote side acknowledges reception of the data, the sender frees the transmit frame buffer 80 and the edge of the TCP window moves to the right. The size of the transmission frame buffer 80 is similar to that of the reception frame, buffer 90, since outstanding data that has not been acknowledged are buffered there, thereby allowing the sender to retransmit in case the receiver on the remote side has not received one or more of the datagrams. Similar to the receive side, this is also a store & forward architecture.
Further limitations and disadvantages of conventional and traditional approaches will become apparent to one of ordinary skill in the art through comparison of such systems with some aspects of the present invention as set forth in the remainder of the present application with reference to the drawings.